MULTIMODAL EMBEDDINGS BEYOND A SINGLE VECTOR

Summary from the TLDR AI newsletter

The Perplexity team launched PPLX-EMBED-V2-LATE, providing multi-vector, multimodal embeddings that enable richer retrieval across text and images. These models retain token-level vectors and utilize a shared embedding space, leading to improved retrieval on benchmarks like ViDoRe(V3). Available in two sizes, 0.6B and 9B, they offer industry-leading performance for various retrieval tasks while maintaining efficiency.